def discriminator(inp):
global reuse_dis, bn_training, blknum
blknum = 0
with tf.variable_scope('discriminator', reuse=reuse_dis):
mod = Model(inp)
mod.set_bn_training(bn_training)
mod.convLayer(7,
16,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) # 64
mod.convLayer(5,
32,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) # 32
mod.SelfAttention(4)
feat = mod.convLayer(5, 64, stride=2, activation=M.PARAM_LRELU) # 16
mod.batch_norm()
mod.convLayer(3,
128,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) # 8
mod.convLayer(3,
128,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True)
adv = mod.convLayer(3, 1)
mod.set_current_layer(feat)
mod.convLayer(3,
128,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True)
block(mod, 128, 1) # 8
mod.convLayer(3,
128,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True)
block(mod, 256, 1) # 4
mod.convLayer(3,
128,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True)
block(mod, 256, 1) # 2
mod.flatten()
age = mod.fcLayer(1)
reuse_dis = True
return adv, age
def discriminator_f(inp, id_num):
global reuse_dis_f, bn_training
with tf.variable_scope('dis_f', reuse=reuse_dis_f):
mod = Model(inp)
mod.set_bn_training(bn_training)
mod.flatten()
feat = mod.get_current_layer()
mod.fcLayer(512, activation=M.PARAM_LRELU, batch_norm=True)
mod.fcLayer(256, activation=M.PARAM_LRELU, batch_norm=True)
adv = mod.fcLayer(1)
mod.set_current_layer(feat)
ip = mod.fcLayer(id_num)
reuse_dis_f = True
return adv, ip
def generator_att(inp):
global reuse_genatt, bn_training, blknum
blknum = 0
with tf.variable_scope('gen_att', reuse=reuse_genatt):
mod = Model(inp)
mod.set_bn_training(bn_training)
mod.convLayer(5,
32,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) #64
block(mod, 64, 1)
mod.convLayer(3,
128,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) #32
block(mod, 128, 1)
# block(mod,256,1)
# block(mod,256,1)
block(mod, 256, 1)
mod.SelfAttention(64, residual=True)
# block(mod,512,1)
block(mod, 256, 1)
# block(mod,256,1)
block(mod, 128, 1)
mod.deconvLayer(3,
64,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) # 64
block(mod, 64, 1)
feat = mod.deconvLayer(5,
64,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) #128
A = mod.convLayer(5, 1, activation=M.PARAM_SIGMOID) #output_attention
mod.set_current_layer(feat)
C = mod.convLayer(5, 3, activation=M.PARAM_TANH)
reuse_genatt = True
return A, C
def generator_att(inp):
global reuse_genatt, bn_training
with tf.variable_scope('gen_att', reuse=reuse_genatt):
mod = Model(inp)
mod.set_bn_training(bn_training)
mod.deconvLayer(3,
512,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) #4
mod.deconvLayer(3,
256,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) #8
mod.deconvLayer(3,
128,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) #16
mod.SelfAttention(32)
mod.deconvLayer(5,
64,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) #32
mod.deconvLayer(5,
32,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) #64
feat = mod.deconvLayer(5,
16,
stride=2,
activation=M.PARAM_LRELU,
batch_norm=True) #128
A = mod.convLayer(5, 3, activation=M.PARAM_SIGMOID) #output_attention
mod.set_current_layer(feat)
C = mod.convLayer(5, 3, activation=M.PARAM_TANH)
reuse_genatt = True
return A, C